Go top
Ph. D. Thesis information

Concentrations in machine learning and energy systems

Stephen J. Lee

Supervised by I.J. Pérez-Arriaga, J.W. Fisher III, J. Taneja, R.J. Stoner, G.C. Verghese

Massachussets Institute of Technology. Cambridge (United States of America)

August 17th, 2023


Economic development requires electric power; however, 770 million people live without access to electricity and 3.5 billion have unreliable connections. Current levels of investment in electricity infrastructure in low- and middle-income countries (LMICs) are significantly below what is needed to achieve global development goals. Energy systems optimization models can enhance electrification planning and resource utilization in LMICs but their effectiveness is constrained by data scarcity pertaining to electricity demand.

First, we evaluate the importance of accurate demand estimation using energy system optimization models and sensitivity analyses. Our findings underscore that accurate demand estimation at high spatial resolution is critical for effective infrastructure planning. To address data gaps in LMICs, we introduce three probabilistic data fusion models for enhanced demand characterizations: the AMPED model for forecasting country-level electricity demand, the BEACON model for estimating building-level electricity access rates, and the LItLDF model for estimating building-level electricity demand.

In this presentation, we highlight the LItLDF model. We train the model using remote sensing and observed metered consumption data. We then use the model to estimate demand in areas lacking consumption data. A significant challenge arises from imprecision in meter geolocation data and ambiguity mapping meters and buildings. We design a probabilistic graphical model (PGM) to characterize relationships between meters and nearby buildings and we incorporate neural network (NN) embeddings to extract information from multimodal building-level features. We use the LItLDF model to estimate electricity demand in Rwanda and Kenya and outline plans to expand the model across Africa.

S.J. Lee (2023), Concentrations in machine learning and energy systems. Massachussets Institute of Technology. Cambridge (United States of America).